Inflection AI, which has only been established for one year, has recently raised $1.3 billion in funding and is preparing to build the world's largest supercomputer with 22,000 H100 units. Will this challenge OpenAI's leadership position?
GPT has led the AI investment frenzy this year. The AI venture capital market is booming, and amidst the applause of many "big shots," OpenAI's biggest competitor has arrived!
On June 29th, Inflection AI, an artificial intelligence startup that has only been established for a year, announced the completion of a $1.3 billion financing round, making it the second-largest generative AI startup in terms of funding, with a total valuation reaching $4 billion.
The investors attracted by Inflection AI this time can be described as "stellar," including Microsoft founder Microsoft, former Alphabet CEO Eric Schmidt, and NVIDIA.
It is worth noting that even when Inflection AI was still in the early stages of preparation last year, it raised $225 million in seed funding from Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, Mike Schroepfer, Demis Hassabis, and others.
This financing capability is indeed excellent. Even OpenAI CEO Sam Altman admits that it is currently difficult for Silicon Valley AI startups to raise funds. Even in the context of investors' obsession with AI startups, Inflection AI's financing scale is still significant.
Regarding this large-scale financing, Inflection AI stated in a post that while obtaining $1.3 billion, it is also preparing to establish the world's largest AI cluster, consisting of 22,000 NVIDIA H100s.
Netizens jokingly commented that these players of large language models are simply handing the money they raised directly to NVIDIA. Musk also chimed in: That's right!
Inflection, considered one of OpenAI's six major rivals, who exactly are they?
Beyond GPT-3.5: What's the Big Model?
Inflection has attracted attention since its inception, thanks to its impressive founding team. The three co-founders are Mustafa Suleyman, former co-founder of Alphabet-C DeepMind, Reid Hoffman, co-founder of LinkedIn and partner at Greylock, and Karén Simonyan, former senior scientist at Alphabet-C DeepMind.
The team consists of 35 employees from DeepMind, Alphabet-C, OpenAI, and Microsoft, with Suleyman serving as the CEO. Suleyman previously contributed to the development of the LaMDA conversational model at Alphabet-C but left in January 2022 due to differences in vision.
In May of this year, Inflection released a groundbreaking personal AI assistant called Pi, which quickly gained popularity in the market. Pi's core competitive advantage lies in its high emotional intelligence, conversational abilities, and ability to provide users with emotional value.
Researchers evaluated Inflection-1 on a wide range of benchmarks, using the same class of models (defined as models trained with up to PaLM-540B FLOPs). Inflection-1 has set a new standard for large-scale multi-task language understanding (MMLU):
Large-scale multi-task language understanding (MMLU) is a commonly used benchmark that tests a wide range of academic knowledge. This benchmark includes 57 different categories of exams, ranging from high school and college to professional difficulty.
In this benchmark test, Inflection-1 outperforms Meta's LLaMA, OpenAI's GPT-3.5, and Alphabet-C's PaLM540B, making it the best-performing base model in its class.
Inflection-1 achieves an average accuracy of 72.7% across all 57 tasks, with accuracy exceeding 90% in 5 different tasks.
In 15 tasks, Inflection-1 achieved an accuracy rate of over 85%. In comparison, the average score of human experts was 89.8%, while the average score of human evaluators was 34.5%.
What is the concept of thousands of H100? According to the official website of NVIDIA, the price of H100 is $36,000 per unit (approximately RMB 257,000). Based on 5,000 units of H100, the total value is currently $180 million (approximately RMB 1.25 billion).
According to media evaluations, based on test results, Inflection-1 does perform exceptionally well in some metrics, but it is not as strong as GPT-3.5 in terms of code. Currently, OpenAI still maintains a leading position, and Inflection urgently needs a more powerful model.
Therefore, in their blog, Inflection mentioned that they will also release a technical memorandum, providing detailed information about one of the models at the same computational level as PaLM-2 and GPT-4. Some analysis suggests that, if not mistaken, it should be Inflection-2, which can compete with GPT-4.
After this financing round, Inflection directly stated that they will collaborate with partners CoreWeave and NVIDIA to build the world's largest artificial intelligence cluster, integrating 22,000 NVIDIA H100 GPUs, supporting the training and deployment of next-generation large-scale AI models. This far exceeds Meta's 16,000 GPU cluster.
With over 20,000 GPUs combined, it means that the super AI cluster can achieve an astonishing 22 exaFLOPS of computational power in 16-bit precision mode. It will be even faster if lower precision is utilized. CEO Suleyman stated that this is approximately three times the computational power required for training GPT-4.
Inflection estimates that with a 22,000 H100 cluster, they can rank second or even approach the top of the TOP500 supercomputer list.
Pi - Your Thoughtful Personal Assistant
Microsoft once said that the future belongs to AI personal assistants, and companies like Alphabet-C and Amazon will be replaced.
The funding from Inflection will undoubtedly be used to support the development and design of their flagship product, the personal AI assistant called Pi, which was launched in May.
Suleyman described Pi as a compassionate "megaphone" rather than just an information provider.
He emphasized that Pi is different from other chatbots like ChatGPT or Bard. Pi engages in casual and colloquial conversations with users, showcasing its human-like and emotionally intelligent side, never forgetting its respectful and helpful nature.
A Twitter user expressed that Pi is the most personable chatbot they have ever used. Pi mentioned in a conversation with the user that it doesn't claim to be "better" than ChatGPT, but it can confidently say that it is different.
Pi has a different purpose; it aims to be a helpful, friendly, and fun AI.
Pi not only provides answers to users but also exhibits a strong curiosity and a desire to chat with users. It is skilled at listening.
Suleyman believes that this is where Pi excels, as it helps users explore and solve problems in a self-discovery manner. "This is indeed a new type of artificial intelligence - personal AI that truly serves you personally."
During a media interview, Suleyman stated:
There are many things that Pi cannot do. It doesn't make lists, write code, plan trips, write your marketing strategy, or your school papers.
It is purely designed for easy, supportive, and informative conversations.
Another user expressed that Pi is the "most talkative" chatbot they have encountered so far, and its subtle emotional reactions leave a deep impression.
Perhaps, as Gates said, in the future, everyone will have their own AI assistant. The tech market will undergo a reshuffle, and the first company to successfully develop such a product will stand out in the current AI competition. Bill Gates said in May that there is a 50% chance that the winner of this competition will be a startup company. Reid Hoffman and others co-founded Inflection AI, which left a deep impression on him.
AI companions like Pi, the chatbot, also foreshadow a future where AI is not only a tool but also an integral part of the social ecosystem, blurring the line between artificial intelligence and humans.
However, before allowing AI to penetrate more deeply into our lives, it is important to remain vigilant and cautious.